Visual Salient Object Detection And Distortion-invariant Optical Correlation Recognition Of Visible Images

There are many applications for visible imaging, such as multimedia processing,machine vision technology, and TV guidance. An effective and accurate detectionand recognition of targets in visible images is critical to the application researchesof visible imaging systems. Generally, targets occupy only a small area of imagesreceived from photoreceptor. A mount of redundancy information not only posesome plausible chanllenges to real time property of system, but also consume a lotof computational resources. Therefore, this dissertation performs the followingresearch works about the salient object detection in cluttered backgrounds anddistortion-invariant recognition with volume holographic correlator of visibleimages.Utilizing the property that natural images are not random but highly structured,we propose an innovative computational mode for salient map construction andsalient objects detection. The model is based on the exploitation of anisotropyproperty of images by means of pixel-wise directional entropy. The generalizedRényi entropy and the one dimentional patch-based discrete cosine transform (DCT)coefficients are selected for this purpose. Calculating the Rényi entropy alongdifferent orientations, four directional entropy maps containing different informationcontent are obtained. Analyzing the power spectrum of entropy map on a log-logscale, we find that the log-log spectra of different images share a similar trend:approximate a straight line which signifies the power law is also appropriate forRényi entropy maps. Furthermore, the low frequency parts which indicate theredundancy span sparsely and linearly, while the high frequency parts whichrepresent the novelty and edges draw together with frequently fluctuations.Therefore, it is possible and reasonable for us to remove the redundancy as well asgenerate saliency maps. Then, the seeded region growing algorithm is employed toextract salient objects from the salient maps. Extensive experiments have beenconducted to verify and evaluate the proposed method.An innovative model for automatically detect salient objects in static colorimages via information content measuring is proposed to produce saliency mapswith well-defined boundaries and effectively suppressed backgrounds. On the basisof searching the local maximum of probability density in joint spatial-range domain,the clusters of pixels that naturally belong together due to a common attribute areobtained. Then we build a dyadic Gaussian pyramid with dynamic scales from theclustered images. The Shannon entropy and the normalized pseudo-Wigner-Ville distribution (PWVD) are selected for the measuring of the information content ofimage components on the clustered images of each scale. The final saliency map isconstructed from Gaussian pyramid by normalizing and summing up informationcontent maps at all scales. In order to demonstrate the validity of the proposedmodel, both qualitative and quantitative experiments have been conducted.We proposed an innovative computational model based on the FIT toautomatically detect visual attention objects in color images without supervisedinteraction and prior knowledge. Simulating human visual attention, three novelfeatures including contrast difference, color gradient, and oriented entropy measureare exploited to describe generic attention objects locally, regionally, and globally.The three features are calculated at multiple scales because of the uncertainty of thesize of attention objects. Based on these feature maps, the final saliency maps areproduced through a nonlinear integration process, rather than simple normalizationand summation. In comparison with the classical approach as well as the other twoproposed methods, the results indicate, overall, that our method achieves relativelybetter detection performance on accuracy and completness.The correlation results of VHC are very sensitive to the distortion such asrotation, scale, and noise of input images. Accordingly we present a novel methodfor VHC to achieve rotation and scale distortion-invariant recognition by using anew mathematical morphology algorithm and SDF technique. In order to suppressthe effect of pattern-dependent behavior on calculated inner product values, originalgray-scale images to be recognized are processed into strong and clean edge maps tomake the correlation patterns of different images similar by using the introducedmorphological algorithm. Since the redundancy is removed from input images, theseprocessed patterns are more appropriate for synthesis of space domain rotation andscale SDF filters to achieve distortion-invariance than gray-scale images. Extensiveexperiments have been conducted on VHC system to verify and evaluate theperformance of proposed method.